Late-interaction / ColBERT-style retrieval
Need multi-vector late interaction retrieval for high-recall RAG where single-vector ANN misses nuanced matches.
At a glance — Oct 2026
Updated 2026-10-27 · forward this page: /jtbds/late-interaction-retrieval#briefScores as of 2026-09-14 (43 days old; Depot marks scores stale after 30 days). The pick is still Depot's current editorial pick; it has not been re-reviewed since. Prices below carry their own dates.
- Depot Pick
- Jina AI · 8.07/10 for this job (curated estimate, high confidence)
Jina historically strong on late interaction / ColBERT-family. Confidence: medium — pin current model card.
- Runner-up
- Voyage AI · 8.39/10
Scores higher than the pick on: Capability fit, Price / performance, Latency / reliability, Maturity / track record.
- When not to use the pick
- Standard embed+rerank (Voyage/Cohere) meets quality bar, or Elastic already stores the corpus.
- Lock-in / migration
- The pick has lower stored lock-in risk than runner-up Voyage AI (3.5 vs 4.2).
- Sourced price (no comparable unit)
Pick: no sourced price in this job's unit (no_matching_sourced_price).
No cheapest SKU (no_comparable_unit_for_job). No price was invented.
- Free option
- No free resource lists this job. A cheap paid SKU is not a free option.
- Failure modes to watch
- Stale retrieval: Store the source URL and the time it was fetched. If a newer canonical page exists, or two stored sources disagree, do not answer from the older one.
- Runaway spend: Set a budget and a tool-call cap per task. Alert when spend jumps. Stop the run at the cap instead of summarizing a partial trace as success.
- Context overflow: If the input exceeds the window, say so and retrieve a narrower slice. Do not answer from a truncated head as if it were the whole source.
- Eval contamination: Keep gold labels out of the agent prompt. Store the harness version beside the score. Do not compare a number across a benchmark revision without saying the revision changed.
- Lost in the middle: Ask the agent to name the passage it used. Spot-check questions whose only support sits in the middle of the context, not only at the edges.
- Policy conflict: If cited sources contradict each other, return the conflict and the URLs. Do not pick a winner silently.
Cite: Agent Provider Depot, Depot Pick for late-interaction-retrieval: jina, scores as of 2026-09-14. Agents: GET /api/decide?jtbd=late-interaction-retrieval
Jina AI
Jina historically strong on late interaction / ColBERT-family. Confidence: medium — pin current model card.
Pricing (directional): As of 2026-10-25 (GET api.jina.ai/v1/models): embeddings/rerank/reader-lm priced at pricing.prompt=0.00000005 → $0.05/1M tokens (jina-embeddings-v3/v4, jina-reranker-v3/v3.5, ReaderLM-v2, clip/colbert). jina-ocr-v1 $0.50/$2.00 per 1M prompt…
Lock-in risk: higher = more lock-in (worse). Overall already inverts this dimension.
Promoted high after pinning official models API token matrix ($0.05/1M). Best as EU-friendly all-in-one embed+rerank+Reader one-liner peer to Cohere/Voyage/Firecrawl. Depot embeddings default stays Cohere; rerank pick stays Voyage.
Quote for Late-interaction / ColBERT-style retrieval
Depot Pick stays Jina AI. Prices below are sourced SKU rows only (as of 2026-10-27). Input and output are not blended. GET /api/quote?jtbd=late-interaction-retrieval
Nothing was compared. no sourced late-interaction unit for this job; chat $/1M was not borrowed pickIsPriceWinner is null.
- jina-embeddings-v3 api_pricing_prompt_field amount unavailable source as of 2026-10-03
- jina-embeddings-v5-text-nano api_pricing_prompt_field amount unavailable source as of 2026-10-03
- jina-ocr-v1 api_pricing_prompt_field amount unavailable source as of 2026-10-03
1 of 4 candidates have no sourced price row. 4 are excluded from the Matching price comparison (none). no sourced late-interaction unit for this job; chat $/1M was not borrowed
Runners-up
Voyage AI
Depot pick for RAG rerank. Keep Cohere as general embeddings default (enterprise + multilingual maturity)…
Cohere
Use for retrieval stacks (embed+rerank), not as your primary agent brain. Quietly excellent at enterprise RAG.
Typesense Cloud
Hybrid search peer when you want Algolia-like DX without per-request fees. Sample $0.03/hr floor pinned…
When NOT to use the top pick
Standard embed+rerank (Voyage/Cohere) meets quality bar, or Elastic already stores the corpus.
Full ranking (score table)
Ranked with this JTBD's weights. Scores are curated estimates unless marked otherwise. Agents: prefer /api/decide then this table for tie-breaks.
| # | Provider | Overall | Fit | $/perf | Latency | DX | As of |
|---|---|---|---|---|---|---|---|
| 1 | Voyage AI | 8.4 | 9.2 | 8.8 | 8.4 | 8.6 | 2026-09-14 |
| 2 | Typesense Cloud | 8.2 | 8.4 | 8.6 | 8.4 | 8.6 | 2026-09-14 |
| 3 | Jina AIDepot Pick | 8.1 | 8.2 | 8.6 | 7.8 | 8.8 | 2026-09-14 |
| 4 | Cohere | 7.9 | 7.8 | 8.0 | 8.4 | 7.8 | 2026-10-19 |
Score methodology for this job
Jina historically strong on late interaction / ColBERT-family. Confidence: medium — pin current model card.
capabilityFit: 32%pricePerformance: 12%latencyReliability: 14%developerDx: 12%privacyPosture: 12%lockInRisk: 8%maturity: 10%